Fix markdown generation crash on LLM special tokens (#4154)

The gpt-tokenizer encode() call rejects special tokens like
<|im_start|> that appear in documentation code examples.
Pass allowedSpecial: 'all' since these are content, not control tokens.

Author: kemister85

scripts/generate-markdown.mjs

@@ -226,7 +226,7 @@ const convertPage = async (htmlPath) => {
 
   const title = extractTitle(dom.window.document);
   const markdown = toMarkdown(articleEl, dom);
-  const tokens = encode(markdown).length;
+  const tokens = encode(markdown, { allowedSpecial: 'all' }).length;
   const content = buildFrontmatter(title, tokens) + markdown + '\n';
   const mdPath = htmlPath.replace(/\.html$/, '.md');